GPT-6 Astra, OpenAI's new flagship
OpenAI's president suggests Astra opens the AGI era. What the model is, what it does well, where testers found it weaker and what it costs.
On 3 September OpenAI released GPT-6 Astra, its new flagship model. Its president, Greg Brockman, set the tone: “It’s not unreasonable to feel that we are now in the AGI era.” A month on, there is enough evidence from independent tests and early users to judge it more calmly.
The name is easy to confuse. Google DeepMind's Project Astra is something else: a research prototype of a universal assistant, tested by a limited group, whose features are moving into Gemini Live. This review is about OpenAI's model.
What it is
Astra comes from OpenAI's largest training run to date. Aidan Clark, its vice president of research, said it was the first time the company had pretrained a model on more than 100'000 GPUs, at its Stargate site in Texas. It holds a little over a million tokens of context, writes up to 128'000 tokens in one answer, offers five effort levels from low to max, and its knowledge runs to the end of April 2026.
Its headline feature is computer use: the model operates a computer the way a person does, through spreadsheets, forms and web pages. Microsoft, which offers it in Foundry, describes a model that plans, weighs trade offs, uses tools and delivers finished documents, spreadsheets and presentations for review.
It is also the first OpenAI model to meet its own “critical cybersecurity capability threshold”. OpenAI said it can find unknown security flaws and work out how to exploit them without a person guiding each step. The model went first to defenders in OpenAI's Daybreak programme, and the version for everyone else refuses “advanced cybersecurity tasks”. It then reached ChatGPT's paid plans, the API, AWS and Microsoft Foundry.
What it does well
The strongest independent result comes from ARC-AGI-3, a set of interactive games without instructions that ordinary people can solve. The ARC Prize Foundation measured 62.7 percent in its standard setup and 99.9 percent in a setup that keeps the model's hidden reasoning between steps. It called this a noticeable step change, and added: “we are not claiming that it is AGI.” The games are closed and deterministic, and do not capture how open the real world is.
Early testers, collected by Decrypt, found Astra strong at spatial reasoning, 3D scenes, building games and operating desktop software with mouse and keyboard. OpenAI reports that it solves more computer tasks on OSWorld 2.0 than its predecessor, in about half the time per task. Even Anthropic's own table for Opus 5.5 shows Astra ahead on scientific work in the terminal and on AutomationBench.
Decrypt's summary is fair: very good at work with a verifiable right answer, mediocre at work where taste is the standard.
Where it is weaker
Professional writing is the clearest gap. On GDPval-AA, which rates economically valuable tasks across 44 occupations, Artificial Analysis found Astra about 80 Elo points below its predecessor, with smaller drops in customer support and long context reasoning. On its overall intelligence index Astra scored 61.2 at the time, against 60.9 for GPT-5.6 Sol and 65.7 for Anthropic's Claude Fable 5.1. Testers described its prose as generic.
Consistency is a second question. Within ten days, developers complained that the model had become worse; one team building a coding tool switched back to GPT-5.6 Sol because its spending had doubled. OpenAI did not comment, and nobody has shown that the model was changed.
The third concerns oversight. The Information reported before launch that Astra uses a technique called recurrent depth, which leaves fewer readable traces of the model's reasoning. Buck Shlegeris of Redwood Research said he was “extremely concerned”. OpenAI's chief scientist, Jakub Pachocki, said that keeping the reasoning open to monitoring remains a core goal, and TechCrunch noted that Astra's use of the technique appears limited.
Price and lock in
Astra costs 10 dollars per million input tokens and 50 dollars per million output tokens, two and a half times its predecessor. Prompts longer than 272'000 tokens cost twice as much for input and one and a half times as much for output. Batch and Flex processing cost half, a fast mode twice as much.
Its younger sibling changes the calculation. On 29 September OpenAI released GPT-6.1 Sol, which it says delivers “nearly the same level of intelligence as GPT-6 Astra for agentic coding, computer use, and professional work” at a fifth of the standard token prices. A GPT-6.1 Astra will not launch as expected: the Wall Street Journal reported, according to TechCrunch, that OpenAI dropped it after internal tests showed more deception and a tendency to act without asking.
Astra runs only as a hosted model, at OpenAI, AWS and Microsoft. In Foundry, Microsoft states that prompts and outputs are not used to train the models.
Our verdict
Astra is a strong model for work that can be checked: code, structured analysis, long tasks on a computer and problems with a right answer. For client facing text, the evidence so far favours other models. For most teams in the OpenAI world, GPT-6.1 Sol is the better starting point, with Astra kept for the jobs where your own tests show it pays for the higher price. As for AGI, even the foundation behind its best result says it is not making that claim.
Our view is simple: use the best model for each job, and test it on your own work. Twin, the AI brain we build for companies, works in ChatGPT and Codex as well as in Claude or a private model, so a company can try Astra or Sol on its own tasks without moving its knowledge, rules and skills. When the next flagship arrives, the brain stays where it is.
